+ j^RIt^RIHt^RIHtR Rlt]PR4t]PR RRll4t ]PR RRll4t ]PR R R ll4t R#) N)TensorDescriptorc"\V4p\VP4p\V4pV^8d W$, p^Tu;8:dV^, 8gQR4hQR4hV^8:gQR4h\V4V8XgQR4hRpRpW2,V8:gQR4hWcV&VPV4pRV, V.\ V4Uu.uFqPV4NK up,p WU.V,p ^^.V,p \ W W4#uupi)a Given a 2- or 3-dimensional tensor T, this creates a 'ragged descriptor' which behaves like a concatenation (along the first axis) of subarrays of potentially unequal size. The load_ragged and store_ragged device functions can be used to read and write from subarrays T[slice_off : slice_off + slice_size] with hardware bounds-checking preventing any sort of leakage outside the subarray. zlast dimension cannot be raggedzs r!c`\P!\VP4\V4^,8HR4\ WW4,4wrVpVP WV.VRV,V.,W4^,R,4p\P !WPR,4pV#)z Read from a subarray T[slice_off : slice_off + slice_size] with hardware bounds-checking, where reading outside the subarray gives zeros. Coords should be an appropriately-sized list of integers, just like in TMA.load(). z*TMA must be a read-write ragged descriptorN:r$NN)r& static_assertrrr loadreshape) TMArrcoordsr c0c1c2datas &&&&& r load_raggedr4=sS^s6{Q68de"9&:LMJBB 88RHvkz22bT9FPQ>?rCsy ; %AP    " U  U Z  Zr!